GitHub Copilot Fusion Matches Near-Opus 5.5 Performance at Half the Cost

Quick Facts

• GitHub Copilot Fusion is now generally available, routing tasks across models from Anthropic and OpenAI including Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol.
• In independent benchmarks by Artificial Analysis, Fusion Max averaged $5.59 per task compared to $11.42 for Claude Code with Opus 5.5 Max, a 51% cost reduction.
• Fusion Max scored within two points of Claude Opus 5.5 Max on the Artificial Analysis Coding Agent Index while using 40% less time per task and roughly half the tokens.

GitHub launched Copilot Fusion at GitHub Universe on October 28, a generally available runtime orchestration system that dynamically selects and coordinates AI models to complete coding tasks rather than relying on any single model throughout a session.

The product grows out of Project HydraFusion, a research preview GitHub announced on September 4, 2026. Where the prior Auto V2 system picked one optimal model per request, HydraFusion and now Copilot Fusion build full execution plans that can invoke multiple models within a single turn.

How It Works

Developers choose one of three effort tiers: Fusion Standard for routine coding, Fusion High for tasks requiring deeper reasoning, and Fusion Max for the most complex work. The system then handles all model selection and coordination behind the scenes.

Copilot Fusion draws from a pool that includes Claude Opus 5.5, Claude Sonnet 5.5, GPT-6 Astra, and GPT-6 Sol. GitHub says the system holds to the same model across a conversation where possible to reuse cached tokens, and only adds model calls when a task warrants escalation.

GitHub Chief Product Officer Mario Rodriguez said the economics come from routing and orchestration choices, not from GitHub subsidizing expensive requests. Each model call is billed at the model's standard rate, and there is no additional charge for Fusion itself.

Benchmark Results

According to independent testing by Artificial Analysis, Copilot Fusion running at Max in the Copilot CLI recorded an overall score of 60 on the Coding Agent Index and an average task cost of $5.59. Claude Code with Opus 5.5 Max scored higher on FrontierCode but cost $11.42 per task on average.

Fusion leads Terminal-Bench 4.0 with a score of 71. On FrontierCode it ranks third at 51, behind Claude Code with Opus 5.5 Max at 56 and Devin Fusion at 54. GitHub's internal DeepSWE benchmark shows Fusion High and Max scoring up to 8.7 points higher than comparable frontier-model efforts and cutting estimated cost by up to 77%.

Fusion Max used approximately 7.2 million tokens per task in the Coding Agent Index, about half the token volume of Claude Opus 5.5 in Claude Code and 30% less than Devin Fusion. The 40% time-per-task reduction is a GitHub-reported figure and is not independently confirmed by Artificial Analysis data.

Business Case

Microsoft EVP of CoreAI Jay Parikh stated that Fusion Max scored within two points of Claude Opus 5.5 Max at half the cost. Microsoft CEO Satya Nadella echoed those numbers in a LinkedIn post, saying Fusion delivers frontier-level quality 40% faster and at up to 77% lower cost in some scenarios.

For engineering teams managing AI coding costs at scale, the gap between $5.59 and $11.42 per task compounds quickly. Hari Pokhrel, a principal software engineer at AT&T, described a common workflow problem Fusion addresses: developers reaching for a powerful model on a hard task and forgetting to switch back for routine work.

The system removes that decision entirely. Rodriguez said Fusion is designed to reduce the burden of model selection while improving the economics of completed work. GitHub's stated goal is to make GitHub the most comprehensive software delivery system available, with developers directing work rather than managing model menus.

Rodriguez said GitHub is evaluating whether to converge HydraFusion into the existing Auto model-selection feature, which would fold multi-model orchestration into a single unified experience for all Copilot users.